ArgueBERT: How To Improve BERT Embeddings for Measuring the Similarity of Arguments

Page created by Terrance Schwartz
 
CONTINUE READING
ArgueBERT: How To Improve BERT Embeddings for Measuring the
                      Similarity of Arguments

                   Maike Behrendt                                Stefan Harmeling
               Heinrich Heine University                      Heinrich Heine University
             maike.behrendt@hhu.de                             harmeling@hhu.de

                      Abstract                            In our work, we focus on exactly these large-
                                                       scale tasks. We want to train embeddings of ar-
    Argumentation is an important tool within hu-      guments in order to measure their similarity, e.g.,
    man interaction, not only in law and politics      to automatically recognize similar user entries in
    but also for discussing issues, expressing and     ongoing discussions in online argumentation sys-
    exchanging opinions and coming to decisions        tems. In this way redundancy can be avoided when
    in our everyday life. Applications for argu-
    mentation often require the measurement of
                                                       collecting arguments. We base our approach on
    the arguments’ similarity, to solve tasks like     Sentence-BERT (SBERT), proposed by Reimers
    clustering, paraphrase identification or summa-    and Gurevych (2019b), which is a bi-encoder, fine-
    rization. In our work, BERT embeddings are         tuning the model’s parameters to place similar sen-
    pre-trained on novel training objectives and af-   tences close to one another in the vector space.
    terwards fine-tuned in a siamese architecture,     This approach yields good results on paraphrase
    similar to Reimers and Gurevych (2019b), to        identification tasks, but evaluating it on an argu-
    measure the similarity of arguments. The ex-
                                                       ment similarity corpus shows a noticeable drop in
    periments conducted in our work show that a
    change in BERT’s pre-training process can im-      performance.
    prove the performance on measuring argument           To improve this method, we propose and evalu-
    similarity.                                        ate three alternative pre-training tasks that replace
                                                       the next sentence prediction (NSP) in BERT’s pre-
1   Introduction                                       training process to optimize SBERT for measuring
                                                       the similarity of arguments. These proposed tasks
Since today it is common to share opinions on so-      are similarity prediction, argument order prediction
cial media to discuss and argue about all kinds        and argument graph edge validation. Being pre-
of topics, the interest of research in the field       trained on these tasks and fine-tuned in a siamese
of artificial intelligence in argumentation is con-    SBERT architecture, we call these models argue-
stantly rising. Tasks like counter-argument re-        BERT throughout this work.
trieval (Wachsmuth et al., 2018), argument cluster-       To examine the models’ applicability in practice,
ing (Reimers et al., 2019a) and identifying the most   we also propose a new evaluation task, which is
prominent arguments in online debates (Boltužić      called similar argument mining (SAM). Solving
and Šnajder, 2015) have been examined and au-         the task of SAM includes recognizing paraphrases
tomated in the past. Many of these tasks involve       (if any are present) in a large set of arguments, e.g.,
measuring the textual similarity of arguments.         when a user enters a new argument to an ongoing
   Transformer-based language models such as           discussion in some form of argumentation system.
the bi-directional encoder representations from           In summary our contributions of this paper are
transformers (BERT) by Devlin et al. (2019) are        the following:
widely used for different natural language process-
                                                         1. We propose and evaluate new pre-training ob-
ing (NLP) tasks. Nevertheless, for large-scale tasks
                                                            jectives for pre-training argument embeddings
like finding the most similar sentence in a collec-
                                                            for measuring their similarity.
tion of sentences, BERT’s cross-encoding approach
is disadvantageous as it creates a huge computa-         2. We propose a novel evaluation task for argu-
tional overhead.                                            mentation systems called SAM.
2   Related Work                                          terms of their similarity. They achieve the best
                                                          results with a fine-tuned BERT model, when incor-
Alternative Pre-Training Objectives The orig-             porating topic knowledge into the network.
inal BERT model uses two different pre-training              In a proximate work Reimers and Gurevych
objectives to train text embeddings that can be used      (2019b) introduce SBERT which serves a base for
for different NLP tasks. Firstly masked language          our work. They train a BERT model in a siamese
modeling (MLM) and secondly next sentence pre-            architecture to produce embeddings of textual input
diction. However, Liu et al. (2019) have shown that       for tasks like semantic similarity prediction. The
BERT’s next sentence prediction is not as effective       model is described in detail in Section 3.1.
as expected and that solely training on the MLM              Dumani et al. (2020) build upon the work of
task can slightly improve the results on downstream       Reimers et al. (2019a) and propose a framework
tasks. Since then there have been attempts to im-         for the retrieval and ranking of arguments, which
prove the pre-training of BERT by replacing the           are both sub-tasks of an argument search engine.
training objectives.                                         Thakur et al. (2020) present an optimized ver-
   Lewis et al. (2020) propose, inter alia, token dele-   sion of SBERT and publish a new argument sim-
tion, text infilling and sentence permutation as alter-   ilarity corpus, which we also use for evaluation
native pre-training tasks. Their experiments show         in this work. They expand the training data for
that the performance of the different pre-training        the SBERT model through data augmentation, us-
objectives highly depends on the NLP task it is           ing the original BERT model for labeling sentence
applied to. Inspired by this we want to explore           pairs.
tasks that perform well on measuring the semantic            To the best of our knowledge there are cur-
similarity of arguments.                                  rently no contextualized embeddings developed
   Lan et al. (2020) propose a sentence ordering          especially for the task of measuring the similarity
task instead of the next sentence prediction, which       of arguments.
is similar to our argument order prediction. They
find that sentence ordering is a more challenging         3     Background
task than predicting if a sentence follows another
sentence. Instead of continuous text, we use dialog       In this section the SBERT (Reimers and Gurevych,
data from argumentation datasets, as we hope to           2019b) architecture, the training procedure and
encode structural features of arguments into our          characteristics are explained in detail.
pre-trained embeddings.
                                                          3.1    SBERT
   Clark et al. (2020) use replaced token detection
instead of MLM, where they do not mask tokens             We use SBERT, proposed by Reimers and
within the sentence, but replace some with alter-         Gurevych (2019b) to fine-tune the BERT models
native tokens that also fit into the sentence. In         pre-trained with our novel proposed pre-training
this way they implement a contrastive learning ap-        tasks.
proach into BERT’s pre-training, by training the             SBERT is a network architecture that fine-tunes
model to differentiate between real sentences and         BERT in a siamese or triplet architecture to cre-
negative samples. Their approach outperforms a            ate embeddings of the input sentences to measure
model pre-trained on MLM on all tasks.                    their similarity. Unlike the original BERT model,
                                                          SBERT is a bi-encoder, which means it processes
Argument Embeddings Embeddings of textual                 each input sentence individually, instead of con-
input that encode semantic and syntactical features       catenating them. The advantage of bi-encoders is
are crucial for NLP tasks. Some research has al-          their efficiency. Cross-encoders like BERT gener-
ready been conducted using the BERT model or its          ate an enormous computational overhead for tasks
embeddings to measure the similarity of arguments.        such as finding the most similar sentence in a large
These are described briefly in the following.             set of sentences, or clustering these sentences.
   Reimers et al. (2019a) use, inter alia, BERT for          By connecting both input sequences, handling
argument classification and clustering as part of an      it as one input, BERT is able to calculate cross-
open-domain argument search. This task involves           sentence attention. Although this approach per-
firstly classification of arguments concerning their      forms well on many tasks, it is not always applica-
topic, and afterwards clustering the arguments in         ble in practice. SBERT is much faster and produces
4     argueBERT
                                                         4.1    Pre-Training
                                                         We propose and evaluate three new tasks, which
                                                         should improve the performance of BERT embed-
                                                         dings on measuring the similarity of arguments.
                                                         The proposed pre-training objectives that are opti-
                                                         mized instead of the next sentence prediction are
                                                         the following:

                                                             1. Similarity prediction: Given a pair of input
                                                                sentences s1 and s2 , predict whether the two
                                                                sentences have the same semantic meaning.
                                                                BERT therefore is pre-trained on the Para-
Figure 1: SBERT architecture for measuring sentence             phrase Adversaries from Word Scrambling
similarity.
                                                                (PAWS) (Zhang et al., 2019) and the Quora
                                                                Question Pairs (QQP)1 dataset.
results that outperform other state-of-the-art embed-        2. Argument order prediction: Given an ar-
ding methods (Reimers and Gurevych, 2019b).                     gumentative dialog consisting of a statement
   To fine-tune the model the authors propose dif-              and an answer to that statement, predict if the
ferent network structures. For regression tasks, e.g.,          given paragraphs p1 and p2 are in the correct
measuring sentence similarity, they calculate the               order. For this task we train BERT on the
cosine similarity of two embeddings u and v, as                 Internet Argument Corpus (IAC) 2.0 (Abbott
shown in Figure 1, and use a mean squared er-                   et al., 2016), which contains argumentative
ror (MSE) loss as objective function. To calculate              dialogues from different online forums. This
the fixed sized sentence embeddings from each in-               task is the same as the sentence ordering ob-
put, a pooling operation is applied to the output               jective from ALBERT (Lan et al., 2020) but
of the BERT model. The authors experiment with                  with argument data.
three different pooling strategies, finding that tak-
                                                             3. Argument graph edge validation: Given
ing the mean of all output vectors works best for
                                                                two arguments a1 and a2 from an argument
their model.
                                                                graph, classify if they are adjacent, thus con-
   In the siamese architecture the weights of the               nected through an edge in the graph. For
models are tied, meaning that they receive the same             this task we use several argument graph
updates. In this way the BERT model is fine-tuned               corpora, taken from http://corpora.aifdb.
to create sentence embeddings that map similar                  org/ for pre-training.
sentences nearby in the vector space.
   In the original paper, the model is fine-tuned on        The pre-training process of argueBERT is the
the SNLI (Bowman et al., 2015) and the Multi-            same as for the original BERT, except that we re-
Genre NLI datasets (Williams et al., 2018) to solve      place the next sentence prediction task. Our novel
multiple semantic textual similarity tasks, which        proposed pre-training objectives are trained as bi-
leads to improved performance in comparison to           nary classification tasks.
other state-of-the-art embedding methods. How-              To compare the new pre-training tasks, we train
ever, evaluating the model on the argument facet         medium sized BERT models with 8 layers and a
similarity (AFS) (Misra et al., 2016) dataset shows      hidden embedding size of 512 (Turc et al., 2019).
a significant drop in accuracy. Different than in        We train the models for a total of 100, 000 training
our work, the authors do not pursue the measure-         steps. To guarantee comparability we also train a
ment of argument similarity in the first place, but      model with the original NSP and MLM objectives
rather use the model for general textual similarity      for 100, 000 steps on the BookCorpus (Zhu et al.,
tasks. The aim of this work is therefore to optimize     2015). To examine if the pre-training tasks also
BERT’s pre-training process to generate argument            1
                                                              https://www.kaggle.com/c/
embeddings that lead to better results on this task.     quora-question-pairs/data
perform on a larger scale, we additionally train a           as being a paraphrase. We calculate the accuracy
BERTBASE model (12 layers, hidden embedding                  and the F1 score of the models on this task.
size 768) on our best performing pre-training task
for 1, 000, 000 steps. All hyperparameters we used           5     Experiments
for pre-training can be found in Table 5 in the Ap-          5.1       Datasets
pendix.
                                                             We use the following datasets for the evaluation of
4.2   Fine-Tuning                                            our embeddings.
For     fine-tuning    argueBERT,       we      use              • The Microsoft Research Paraphrase Corpus2
SBERT (Reimers and Gurevych, 2019b).                               (MSRP) (Dolan and Brockett, 2005), which
The model fine-tunes the weights of the pre-trained                includes 5, 801 sentence pairs for paraphrase
BERT model in a siamese architecture, such that                    identification with binary labeling (0: “no
the distance between embeddings of similar input                   paraphrase”, 1: “paraphrase”), automati-
sentences is minimized in the corresponding vector                 cally extracted from online news clusters.
space. Therefore, ŷ is calculated as the cosine
similarity between two input embeddings u and v                  • The Argument Facet Similarity Dataset3
and then the MSE loss                                              (AFS) (Misra et al., 2016), consisting of
                            n
                                                                   6, 000 argument pairs taken from the Inter-
                        1X                                         net Argument Corpus on three controversial
              MSE =       (yi − ŷi )2                (1)
                        n                                          topics (death penalty, gay marriage and gun
                           i=1
                                                                   control), annotated with an argument facet
is optimized. Here n is the batch size and y the true              similarity score from 0 (“different topic”) to
label. We fine-tune each model on every evaluation                 5 (“completely equivalent”).
dataset for a total of five epochs with a batch size of
16 and a learning rate of 2e-5. All hyperparameters              • The BWS Argument Similarity Dataset4
used for fine-tuning can be found in Table 6 in the                (BWS) (Thakur et al., 2020), which contains
Appendix.                                                          3, 400 annotated argument pairs on 8 contro-
                                                                   versial topics from a dataset collected from
4.3   Similar Argument Mining (SAM)                                different web sources by Stab et al. (2018b).
The main idea of proposing argueBERT as an im-                     Labeled via crowd-sourcing with similarity
proved version of SBERT on measuring argument                      scores between 0 and 1.
similarity is in particular to use it for identifying            • The UKP Argument Aspect Similarity Cor-
and mining similar arguments in online argumenta-                  pus5 (UKP) (Reimers et al., 2019a) with a
tion systems. In order to evaluate language models                 total of 3, 595 argument pairs, annotated with
for this purpose, we propose a new evaluation task                 four different labels “Different topic/ can’t de-
which we call SAM. It is defined as follows.                       cide”, “no similarity”, “some similarity” and
Task definition. Given a query argument q,                        “high similarity” on a total of 28 topics, which
match the argument against all arguments of an                     have been identified as arguments by the Ar-
existing set S = {a1 , a2 , . . . , an } \ {q} to predict,         gumenText system (Stab et al., 2018a).
if S contains one or more paraphrased versions of
                                                             As baselines we use (i) a medium sized SBERT,
q and find the paraphrased sentences in the set.
                                                             pre-trained with the standard BERT pre-training
                                                             procedure, fine-tuned in a siamese architecture, and
   For the evaluation on SAM, the model is given             (ii) average word2vec6 (Mikolov et al., 2013) vec-
a set of arguments of which some are paraphrased
                                                                   2
argument pairs and some are unpaired arguments                    https://www.microsoft.com/en-us/
                                                             download/details.aspx?id=52398
that are not considered equivalent to any other ar-             3
                                                                  https://nlds.soe.ucsc.edu/node/44
gument in the set. The model then encodes all                   4
                                                                  https://tudatalib.ulb.tu-darmstadt.
arguments into vector representations and calcu-             de/handle/tudatalib/2496
                                                                5
lates the pairwise cosine similarities. If the highest            https://tudatalib.ulb.tu-darmstadt.
                                                             de/handle/tudatalib/1998
measured similarity score for an argument exceeds               6
                                                                  https://code.google.com/archive/p/
a pre-defined threshold, the argument is classified          word2vec/
MSRP           UKP            AFS
                  Model                            r      ρ       r     ρ        r     ρ
                  average word2vec               18.17 17.96    22.29 17.44    11.25  5.22
                  SBERT                          47.12 44.54    32.04 30.89    38.02 35.92
                  argueBERT sim. pred. (ours)    48.33 46.34    35.33 34.77    37.57 35.83
                  argueBERT order pred. (ours)   45.08 43.15    28.41 28.11    38.25 36.80
                  argueBERT edge val. (ours)     40.88 40.03    28.36 26.64    36.89 34.04

 Table 1: Pearson’s correlation r and Spearman’s rank correlation ρ × 100 on the MSRP, UKP and AFS corpora.

tors with vector-size 300, pre-trained on part of the    6     Results
Google News dataset.
                                                         First of all, we evaluate how well our models can
   To be able to fine-tune the models, the discrete      predict the similarity of a given argument pair by
labels of the AFS and UKP corpus are transformed         calculating the cosine similarity between the two
into similarity scores between 0 and 1. The labels       embeddings. Table 1 shows the Pearson correlation
of the AFS corpus, which range from 0 to 5, are          r and Spearman’s rank correlation ρ on this task
normalized by dividing it through the maximum            for the MSRP, AFS and UKP datasets.
value of 5. For the UKP corpus, the labels ”dif-            On the MSRP dataset, the model pre-trained with
ferent topic/ can’t decide” and ”no similarity” are      a similarity prediction objective performs slightly
assigned the value 0, ”some similarity” is trans-        better than the baseline that is trained with the next
lated into a similarity score of 0.5 and for all pairs   sentence prediction objective. The argueBERT or-
with label ”high similarity” we assign a similarity      der prediction model only performs a little worse
score of 1. The labels for the MSRP corpus remain        on this dataset, than the next sentence prediction
unchanged.                                               model, while the model trained on edge validation
                                                         can not compete with the aforementioned models.
  We perform two different evaluations. Firstly
                                                            On the UKP dataset the performance increase
on the task of similarity prediction. Therefore we
                                                         by the model that used the similarity prediction
evaluate the models by calculating the Pearson’s
                                                         objective for pre-training is even more significant.
and Spearman’s rank correlation for the predicted
                                                         It outperforms the traditionally pre-trained SBERT
cosine similarities. Secondly we calculate the ac-
                                                         model by 3 points for Pearson correlation and al-
curacy and F1 score on the novel proposed task of
                                                         most 4 points for the Spearman rank correlation.
SAM.
                                                            Surprisingly, the order prediction model is able
   For the AFS corpus, which contains arguments          to outperform the similarity prediction task on the
for three different controversial topics, we use the     AFS corpus. But it has to be noticed that there is
same cross-topic evaulation strategy as suggested        not much difference in the performance of all mod-
by Reimers and Gurevych (2019b). The models are          els on this dataset. Only the averaged word2vec
fine-tuned on two of the three topics and evaluated      vectors perform notably worse than all other evalu-
on the third one, taking the average of all possible     ated models.
cross-topic scenarios as overall model performance          Out of all evaluated datasets, the recently pub-
score.                                                   lished BWS corpus is the only one whose simi-
                                                         larity values are quantified on a continuous scale.
   The UKP corpus, including arguments on 28
                                                         Table 2 shows the evaluation results for all models
different topics, is evaluated with a 4-fold cross-
                                                         for three different distance measures. We chose the
topic validation as done by Reimers et al. (2019a).
                                                         cosine similarity as default distance measure for
Out of the 28 topics, 21 are chosen for fine-tuning
                                                         evaluation. But in the case of the BWS corpus it
the model and 7 are used as test set. The evaluation
                                                         is striking that both Manhattan and Euclidean dis-
result is the averaged result from all folds.
                                                         tance result in a higher Pearson correlation as well
   The BWS argument similarity dataset incorpo-          as Spearman rank correlation. The embeddings
rates 8 different controversial topics. For evaluation   of argueBERT pre-trained with a similarity predic-
we fine-tune the models on a fixed subset (T1 − T5 ),    tion objective achieve the highest correlation for
validate them on another unseen topic (T6 ) and use      all distance measures. The model outperforms the
the remaining two topics as test set (T7 − T8 ), as      SBERT model by 4 points. The argument order pre-
suggested by Thakur et al. (2020).                       diction model also performs better than the model
Cosine        Manhattan      Euclidean
                 Model                             r        ρ       r     ρ       r       ρ
                 average word2vec                8.98     3.46    41.67 43.61   41.73 43.54
                 SBERT                           38.55 38.72      42.33 42.02   42.35 42.09
                 argueBERT sim. pred. (ours)     43.44 43.76      46.84 46.94   46.56 46.70
                 argueBERT order pred. (ours)    38.97 38.02      44.20 43.39   44.14 43.30
                 argueBERT edge val. (ours)      33.72 33.22      39.49 38.79   39.74 39.17

Table 2: Pearson’s correlation r and Spearman’s rank correlation ρ ×100 on the BWS Argument Similarity
Dataset (Thakur et al., 2020) for three different distance measures.

       Model                               ρ                 Model                             Acc.     F1
       average word2vec                  43.54               average word2vec                  35.49   45.68
       SBERTBASE (Thakur et al., 2020)   58.04               SBERT                             44.92   52.54
       argueBERTBASE sim. pred. (ours)   62.44               argueBERT sim. pred. (ours)       64.09   69.80
       BERTBASE (Thakur et al., 2020)    65.06               argueBERT order pred. (ours)      38.14   46.81
                                                             argueBERT edge val. (ours)        48.10   49.08
Table 3: Spearman’s rank correlation ρ ×100 on the           SBERTBASE                         66.88   71.45
BWS argument similarity dataset.                             argueBERTBASE sim. pred. (ours)   65.92   70.76

                                                         Table 4: Accuracy and F1 score on SAM for the MSRP
                                                         corpus for a threshold of 0.8.
pre-trained with next sentence prediction, only the
edge validation argueBERT model does not lead to
an improvement and even performs worse than the          argueBERTBASE model does not perform as well
word2vec baseline approach for both Manhattan            as the SBERT model on this dataset.
and Euclidean distance measures.                            The remaining argument similarity datasets were
   To see how well the pre-training works for larger     found to be unsuitable for the task of SAM as they
models, we also trained an argueBERTBASE model           do not only contain dedicated paraphrased argu-
on the task of similarity prediction for 1, 000, 000     ment pairs, but rather present all increments of sim-
training steps on the PAWS and QQP datasets. The         ilarity. This means that very similar arguments are
evaluation results for the BWS dataset are shown         not necessarily matched as argument pairs in the
in Table 3. For comparison we also list the evalua-      data. Therefore, for future research new datasets
tion result of the standard BERTBASE model on            that suit the task of SAM are required.
this dataset. Even though argueBERTBASE was
                                                         7       Discussion
trained on a comparably small dataset, it outper-
forms SBERT on the BWS argument similarity               Our conducted experiments show that the new pro-
prediction task and almost reaches the level of the      posed pre-training tasks are able to improve the
BERTBASE cross-encoder.                                  SBERT embeddings on argument similarity mea-
   Lastly, Table 4 shows the results on the MSRP         surement, compared to the next sentence prediction
dataset on the task of SAM for both the small and        objective. Nevertheless, our presented approach
large pre-trained models. The small argueBERT            has some limitations that should be addressed in
model, pre-trained with the similarity prediction        the following.
objective, by far achieves the highest accuracy, as         First of all, the proposed models were pre-trained
well as the highest F1 value for a threshold of 0.8.     and fine-tuned on a single GPU. Due to the lim-
This reflects the evaluation results of the sentence     ited resources, a BERT model in medium size was
embeddings on this dataset, showing that the simi-       chosen as basis for all pre-trained models. The
larity prediction argueBERT model is able to recog-      models were trained only for a total of 100, 000
nize paraphrases in the dataset quite well. The sec-     training steps, which is just a small fraction of the
ond best performing models, which are the argue-         conducted training of the original BERT model.
BERT model trained on the task of edge validation        The achieved results have to be regarded as com-
and the baseline, trained on next sentence predic-       parative values on how much an adaptation of the
tion, are almost more than 16 points behind. This        pre-training process can improve the performance.
shows the great potential of incorporating similar-      However, training a larger model for 1, 000, 000
ity prediction in the pre-training process of BERT.      steps on the task of similarity prediction indicates
Looking at the results for the larger models, the        that the adapted pre-training also works for larger
models and is able to compete with a pre-trained        SAM can be used to evaluate models on the abil-
cross-encoder.                                          ity to identify paraphrases from a large collection
   Another point is that the corpora we used for pre-   of sentences. Fields of application are, for exam-
training have quite different characteristics. The      ple, online argumentation tools, where users can
IAC (Abbott et al., 2016) for example consists of       interchange arguments on certain topics. Newly
posts from different online forums. The used lan-       added arguments can be compared to existing posts
guage is colloquial and the posts strongly vary in      and duplicate, paraphrased entries can be avoided.
length and linguistic quality. The same applies to      A trained model that is good at measuring argu-
the QQP corpus. In contrast, the PAWS dataset           ment similarity is also advantageous for tasks like
consists of paraphrases extracted from Wikipedia        argument mining and argument clustering.
articles, implying a formal language without mis-
spellings. Training models on informal datasets
can be advantageous, depending on the application       References
of the trained model. In our case the differences of    Rob Abbott, Brian Ecker, Pranav Anand, and Marilyn
the used datasets rather constitute a disadvantage,       Walker. 2016. Internet argument corpus 2.0: An
as it may affect the comparability of the resulting       sql schema for dialogic social media and the cor-
                                                          pora to go with it. In Proceedings of the Tenth In-
models.                                                   ternational Conference on Language Resources and
   Additionally to having different characteristics,      Evaluation (LREC’16), pages 4445–4452, Portorož,
the few available datasets on paraphrase identifi-        Slovenia. European Language Resources Associa-
cation, argument similarity and also the argument         tion (ELRA).
graph corpora are relatively small, compared to         Filip Boltužić and Jan Šnajder. 2015. Identifying
the corpora the original BERT model is trained             prominent arguments in online debates using seman-
on. For the task of argument similarity prediction         tic textual similarity. In Proceedings of the 2nd
only the recently published BWS corpus (Thakur             Workshop on Argumentation Mining, pages 110–
                                                           115, Denver, CO. Association for Computational
et al., 2020) includes argument pairs annotated with       Linguistics.
continuous scaled similarity scores. It can be said
that there is still a lack of high-quality annotated    Samuel R. Bowman, Gabor Angeli, Christopher Potts,
argumentation corpora for this task.                      and Christopher D. Manning. 2015. A large anno-
                                                          tated corpus for learning natural language inference.
                                                          In Proceedings of the 2015 Conference on Empiri-
8   Conclusion                                            cal Methods in Natural Language Processing, pages
                                                          632–642, Lisbon, Portugal. Association for Compu-
In our work, we proposed and evaluated different          tational Linguistics.
pre-training tasks to improve the performance of
SBERT embeddings on the task of argument sim-           Kevin Clark, Minh-Thang Luong, Quoc V. Le, and
                                                          Christopher D. Manning. 2020.         Electra: Pre-
ilarity measurement. We call the new pre-trained          training text encoders as discriminators rather than
model variants argueBERT. Evaluation of the mod-          generators. In International Conference on Learn-
els shows that adapting the pre-training process of       ing Representations.
BERT has an impact on the resulting embeddings
                                                        Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
and can improve the models’ results. ArgueBERT             Kristina Toutanova. 2019. BERT: Pre-training of
trained with a similarity prediction objective led to      deep bidirectional transformers for language under-
a performance improvement up to 5 points Spear-            standing. In Proceedings of the 2019 Conference
man’s rank correlation on the evaluated BWS ar-            of the North American Chapter of the Association
                                                           for Computational Linguistics: Human Language
gument similarity corpus, compared to the model           Technologies, Volume 1 (Long and Short Papers),
trained with the classic NSP pre-training task and         pages 4171–4186, Minneapolis, Minnesota. Associ-
also showed the best results on our new proposed           ation for Computational Linguistics.
evaluation task SAM on the MSRP corpus.
                                                        William B. Dolan and Chris Brockett. 2005. Automati-
   A larger argueBERTBASE pre-trained with the            cally constructing a corpus of sentential paraphrases.
similarity prediction task could improve the eval-        In Proceedings of the Third International Workshop
uated embeddings compared to SBERT and al-                on Paraphrasing (IWP2005).
most reaches the results of the cross-encoding
                                                        Lorik Dumani, Patrick J. Neumann, and Ralf Schenkel.
BERTBASE model.                                           2020. A framework for argument retrieval - rank-
   For future research, the new proposed task of          ing argument clusters by frequency and specificity.
In Advances in Information Retrieval - 42nd Euro-          heterogeneous sources. In Proceedings of the 2018
  pean Conference on IR Research, ECIR 2020, Lis-            Conference of the North American Chapter of the
  bon, Portugal, April 14-17, 2020, Proceedings, Part        Association for Computational Linguistics: Demon-
  I, volume 12035 of Lecture Notes in Computer Sci-          strations, pages 21–25, New Orleans, Louisiana. As-
  ence, pages 431–445. Springer.                             sociation for Computational Linguistics.

Zhenzhong Lan, Mingda Chen, Sebastian Goodman,             Christian Stab, Tristan Miller, Benjamin Schiller,
  Kevin Gimpel, Piyush Sharma, and Radu Soricut.             Pranav Rai, and Iryna Gurevych. 2018b. Cross-
  2020. Albert: A lite bert for self-supervised learning     topic argument mining from heterogeneous sources.
  of language representations. In International Con-         In Proceedings of the 2018 Conference on Em-
  ference on Learning Representations.                       pirical Methods in Natural Language Processing,
                                                             pages 3664–3674, Brussels, Belgium. Association
Mike Lewis, Yinhan Liu, Naman Goyal, Mar-                    for Computational Linguistics.
  jan Ghazvininejad, Abdelrahman Mohamed, Omer
  Levy, Veselin Stoyanov, and Luke Zettlemoyer.            Nandan Thakur, Nils Reimers, Johannes Daxenberger,
  2020. BART: Denoising sequence-to-sequence pre-            and Iryna Gurevych. 2020. Augmented sbert: Data
  training for natural language generation, translation,     augmentation method for improving bi-encoders for
  and comprehension. In Proceedings of the 58th An-          pairwise sentence scoring tasks. arXiv preprint
  nual Meeting of the Association for Computational          arXiv:2010.08240.
  Linguistics, pages 7871–7880, Online. Association
  for Computational Linguistics.                           Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina
                                                              Toutanova. 2019. Well-read students learn better:
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-           On the importance of pre-training compact models.
  dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,               arXiv preprint arXiv:1908.08962v2.
  Luke Zettlemoyer, and Veselin Stoyanov. 2019.
  Roberta: A robustly optimized bert pretraining ap-       Henning Wachsmuth, Shahbaz Syed, and Benno Stein.
  proach.                                                    2018. Retrieval of the best counterargument with-
                                                             out prior topic knowledge. In Proceedings of the
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Cor-           56th Annual Meeting of the Association for Com-
  rado, and Jeffrey Dean. 2013. Distributed represen-        putational Linguistics (Volume 1: Long Papers),
  tations of words and phrases and their composition-        pages 241–251, Melbourne, Australia. Association
  ality. In Proceedings of the 26th International Con-       for Computational Linguistics.
  ference on Neural Information Processing Systems
  - Volume 2, NIPS’13, page 3111–3119, Red Hook,           Adina Williams, Nikita Nangia, and Samuel Bowman.
  NY, USA. Curran Associates Inc.                            2018. A broad-coverage challenge corpus for sen-
                                                             tence understanding through inference. In Proceed-
Amita Misra, Brian Ecker, and Marilyn Walker. 2016.          ings of the 2018 Conference of the North American
 Measuring the similarity of sentential arguments in         Chapter of the Association for Computational Lin-
 dialogue. In Proceedings of the 17th Annual Meet-           guistics: Human Language Technologies, Volume
 ing of the Special Interest Group on Discourse and          1 (Long Papers), pages 1112–1122, New Orleans,
 Dialogue, pages 276–287, Los Angeles. Association           Louisiana. Association for Computational Linguis-
 for Computational Linguistics.                              tics.

Nils Reimers and Iryna Gurevych. 2019b. Sentence-          Yuan Zhang, Jason Baldridge, and Luheng He. 2019.
  BERT: Sentence embeddings using Siamese BERT-              PAWS: paraphrase adversaries from word scram-
  networks. In Proceedings of the 2019 Conference on         bling. In Proceedings of the 2019 Conference of the
  Empirical Methods in Natural Language Processing           North American Chapter of the Association for Com-
  and the 9th International Joint Conference on Natu-        putational Linguistics: Human Language Technolo-
  ral Language Processing (EMNLP-IJCNLP), pages              gies, NAACL-HLT 2019, Minneapolis, MN, USA,
  3982–3992, Hong Kong, China. Association for               June 2-7, 2019, Volume 1 (Long and Short Papers),
  Computational Linguistics.                                 pages 1298–1308. Association for Computational
                                                             Linguistics.
Nils Reimers, Benjamin Schiller, Tilman Beck, Jo-
  hannes Daxenberger, Christian Stab, and Iryna            Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhut-
  Gurevych. 2019a. Classification and clustering of ar-      dinov, Raquel Urtasun, Antonio Torralba, and Sanja
  guments with contextualized word embeddings. In            Fidler. 2015. Aligning books and movies: Towards
  Proceedings of the 57th Annual Meeting of the As-          story-like visual explanations by watching movies
  sociation for Computational Linguistics, pages 567–        and reading books. In Proceedings of the 2015
  578, Florence, Italy. Association for Computational        IEEE International Conference on Computer Vision
  Linguistics.                                               (ICCV), ICCV ’15, page 19–27, USA. IEEE Com-
                                                             puter Society.
Christian Stab, Johannes Daxenberger, Chris Stahlhut,
  Tristan Miller, Benjamin Schiller, Christopher
  Tauchmann, Steffen Eger, and Iryna Gurevych.
  2018a. ArgumenText: Searching for arguments in
A    Appendix
Pre-Training and fine-tuning settings
Table 5 shows the used settings for pre-training all
proposed BERT models in this work.

    BERT model                  BERT medium uncased,
                                BERT base uncased
    learning rate               1e-4, 2e-5
    do lower case               True
    max seq length              128
    max predictions per seq     5
    masked lm prob              0.15
    random seed                 12345
    dupe factor                 10

  Table 5: Settings for creating the pre-training data.

   Table 6 shows the settings for fine-tuning SBERT
on the evaluated datasets, using the sentence-
transformers library7 published by the UKPLab
on GitHub.
         learning rate        2e-5
         train batch size     16
         num epochs           5
         optimizer class      transformers.AdamW
         weight decay         0.01

           Table 6: Settings for fine-tuning.

   7
     https://github.com/UKPLab/
sentence-transformers
You can also read